Warning
Bridlewise is an experimental developer preview. It is not security-audited, not production-ready, and must not be used with sensitive, proprietary, regulated or production codebases.
There are no supported production versions. main is the only development
branch, and every release is a preview.
Do not open a public issue with exploit details, credentials or private data.
Use GitHub's private vulnerability reporting:
https://github.com/LLM-4-People/bridlewise-testing/security/advisories/new
Include the affected revision, a minimal reproduction, the impact and any suggested fix. A confirmed report is handled in a private security advisory.
There is no incident-response team and no disclosure SLA. Reports are handled on a best-effort basis.
- Loopback API binding and a
Hostcheck. - HttpOnly sessions, a CSRF header on unsafe methods, scrypt password hashes and hashed session tokens.
- AES-256-GCM field encryption at rest, with keys outside the checkout.
- Redaction at the model-egress choke point and before user display.
- A governed tool surface with typed refusals.
- Landlock confinement for spawned commands and stdio MCP servers on Linux.
- An append-only, hash-chained timeline.
- An audit, penetration test or formal threat-model review.
- Multi-tenant hardening or a production deployment model.
- Network confinement.
- Complete secret or PII detection.
- A guarantee that every destructive action is reversible.
- Support or an SLA.
Pull-request CI runs on the pull_request event only, with a read-only token
and no secrets, for every author alike; a first-time contributor's first run
waits for a maintainer's approval. No workflow runs contributor code with write
access or secrets.
- Never commit provider keys, database keys,
.envfiles, tokens, private keys or local databases. - Keep the database key home outside the repository.
- Redact logs, absolute paths, project names and personal data before opening an issue.
- If a credential was exposed, revoke and rotate it first; rewriting history does not undo exposure.
Use a disposable copy. Assume data loss is possible. Inspect the run record before trusting an outcome. Do not point the harness at a codebase you cannot afford to lose or disclose.